Surveillance Video Re-Identification with Robustness to Occlusion

‍ ‍Traditional identification technologies require pre-recorded information from target personals， while failing to consider any visual obstructions in the identification process， resulting in its unsatisfactory performance in surveillance-video-based re-identification scenarios， especially for public spaces. Most existing person re-identification approaches examine appearance features such as clothing and decoration， which are prone to change in time and space， and thus are unreliable for long-term tracking. An effective and reliable approach for long-term re-identification is to utilize stable biometric features such as facial features. However， with occlusion， low resolution， lack of illumination， and perspective gestures exhibited in surveillance videos， traditional facial recognition methods that are excellent for image recognition cannot perform well. To address these issues， this paper proposed a deep-learning-based face re-identification algorithm. The algorithm combined an attention mechanism with a mask dictionary to dynamically and appropriately assign weights to video frame features， thereby reducing the effect of occlusion and effectively improving the re-identification accuracy. Extensive experiments demonstrated that the proposed method was able to achieve a rank-1 accuracy of up to 95.2% on the cox dataset， and 73.0% on the same dataset with synthetic occlusion. These results comfirm the superior performance of the proposed algorithm compared to state-of-the-art re-identification algorithms.